Abstract:To address missed detections of occluded pedestrians and distant small targets, as well as difficulties in accurately localizing the boundaries of overlapping targets in dense crowds, we propose SDT-YOLO, a detection algorithm based on YOLOv11n. A Stagewise Fused and Reparameterized Backbone (SFR-Backbone) extracts locally visible cues and integrates them with high-level semantic information to improve the feature representation of occluded targets. A Dynamic Sampling and Weighted Fusion Path Aggregation Network (DSWF-PAN) adjusts sampling positions through dynamic upsampling and performs multiscale feature fusion using spatially adaptive weighting, strengthening the integration of spatial details and semantic information for small targets. A Task-Conditioned Context Refinement Detection Head (TCR-Detect) selects neighborhood information according to the distinct requirements of classification and regression, supplementing features for pedestrian recognition and boundary localization. On the CrowdHuman dataset, the model achieves a recall of 75.71% and an mAP@0.5:0.95 of 54.31%, exceeding the baseline by 4.19 and 2.74 percentage points, respectively. On the WiderPerson dataset, these two metrics improve by 2.25 and 1.50 percentage points, respectively. The results show that the proposed algorithm reduces missed pedestrian detections and improves mean average precision in dense crowd scenes.